Objective: This project aims to improve reliability and safety of space habitats by developing robust fault detection systems that can operate independently of ground control. The project leverage machine learning algorithms and model selection techniques to identify and classify faults in the habitat systems.

Contribution: I worked on updating a fault detection system for CO2 sensors in the habitat and dockerizing project files to be showen at the final project demonstration. This process taught me the importance of reproducibility and the benefits of containerization in software development. Additionally I was able to leverage my unique background exploring anomaly detection techniques and apply it to a fault detection space

Skills: Python, Docker, Fault Detection, Machine Learning

Objective: Real-time condition monitoring is crucial for infrastructure health, but indirect structural health monitoring (SHM) faces challenges like changing conditions and anomaly detection. Existing machine learning methods lack adaptability to new environments. This study proposes unsupervised anomaly detection using sensor data from diverse train conditions, employing techniques like PCA, ICA, and semantic segmentation for improved monitoring.

Contribution: I led this research initiative, developing a custom data acquisition system that incorporatedI designed and built the data acquisition system, comprising of circuitry to interface the accelerometers, GPS, and regulated power sources while writing custom software to collect and process the data. During deployment, we could not “walk to track” to identify portions of broken rail due to safety concerns. This dilemma posed a challenge as our previous model relied on supervised classification models. I was inspired by unsupervised learning and anomaly detection literature, which led me to develop a stacked autoencoder ensemble that took the short-time Fourier transform as an input and aimed to separate “normal” regions of rail from “abnormal” sections using the reconstruction error as the metric for separation. In a surprising turn of events, we also discovered that the track was free of damaged rail. However, the model detected several locations later confirmed to contain broken cross-ties. I connected with the owner of RailPulse through a fortuitous conversation with my GIS professor, sparking a deal to run remote broken rail testing at his facilities. In preparation for deployment at both facilities, I duplicated the data acquisition system and added vision capabilities that tested a multi-modal classification pipeline that combines vision, acceleration, and GPS data. a novel unsupervised anomaly detection algorithm for real-time broken rail detection in in-service locomotives. The study involved data collection, preprocessing, and model training, culminating in a robust algorithm that outperformed existing methods.

Skills: Python, Unsupervised Machine Learning, Data Acquisition, Data Management, Signal Processing, Field Testing

Objective: In recent research, focus has shifted towards detecting infrastructure damage using onboard acceleration signals, aiming for real-time track damage detection. However, a robust anomaly detection algorithm is lacking for rail crack detection. This study utilizes lab-scale track simulations to evaluate techniques, finding principal component analysis most effective. Meanwhile, there's an urgent need for consistent and automated rail crack detection methodologies, addressed through an open-source lab-scale rail testbed framework.

Contribution: This project consisted of developing a lab-scale track-train model that could accurately simulate vibrations at the interface. I led the model development, collaborating with department staff and industry partners at Wabtec Corp. Working in tandem with a colleague, we built the lab model and identified the impacts of dimensionality reduction techniques (e.g., PCA, ICA, autoencoders, etc.) on support vector machine classification using experimental data: finding a 95% classification accuracy employing PCA.

Skills: Python, Data Acquisition, Signal Processing, Machine Learning, Open Source, Hardware Development

Objective: This project first introduced me to edge computing hardware deployments. The project aimed to optimize building energy consumption by developing a real-time occupancy detection system using "privacy-preserving" Kinect sensors. The project involved deploying a machine learning model to detect and track occupants in a room, interface with a sensor network via ssh, and coordinate with facilities management staff.

Contribution: I worked with department and university-wide staff to take down and redeploy odroids and intel sticks connected to the Microsoft Kinect sensors. I also worked on updating the systems to run pre-trained machine learning classification models that tracked people entering and leaving rooms.

Skills: Edge Computing, Energy Optimization, Machine Learning, Deployment, CLI

Objective: This study investigates the adverse effects of locomotive-induced repeated loading on railway aggregates, leading to aggregate spoiling. Focusing on mitigating this issue, the research evaluates various rail tie padding materials to reduce stress on the aggregate. By testing different padding types, the study aims to identify solutions that effectively minimize aggregate spoiling, thereby enhancing railway durability and performance under locomotive loading conditions.

Contribution: Developed a custom MATLAB pipeline that incorporated digital filtering and peak detection algorithms to identify load peaks when analyzing under rail-tie padding material. Additionally, I had to overcome real-world challenges (e.g., faulty and broken pressure sensors) to ensure accurate data collection and analysis.

Skills: MATLAB, Data Analysis, Signal Processing, Peak Detection

Objective: This study addresses soil movement optimization in space-constrained environments like Singapore through earthwork optimization techniques. By integrating minimal spanning trees and linear programming, the research explores strategies to minimize soil movement while maximizing construction efficiency. The approach unlocks potential cut and fill patterns, optimizing earthwork operations to reduce soil disturbance. Implementing such methods could significantly enhance construction processes in densely populated areas, ensuring sustainable development and resource utilization.

Contribution: This was my first exposure to reseach and I was tasked with exploring computational tools.

Skills: Python, Simulation, Optimization, Linear Programming, Data Analysis